When DeepSeek's inference model shocked the global AI community, a Chinese unicorn valued at 20 billion RMB was quietly sharpening its knives, ready to launch a direct challenge to this newcomer with a training cost of only $530,000 and a disruptive architectural design.
On the 17th, AI startup MiniMax released its first inference model M1, which, according to benchmark evaluations, outperformed domestic closed-source models and approached the most advanced overseas models, with some tasks surpassing DeepSeek, Alibaba, ByteDance, as well as the latest and strongest open and closed-source models from OpenAI, Google, and Anthropic.
The core of this competition lies not only in performance but also in efficiency— compared to DeepSeek R1, M1 consumes less than 50% of the computing power when generating 64K tokens; at 100K tokens, it is only 25%.
MiniMax stated that the entire reinforcement learning process for M1 used 512 NVIDIA H800 GPUs trained for three weeks, with a rental cost of $537,400 (approximately 3.8 million RMB). This cost control &#34;is an order of magnitude less than initially expected.&#34; MiniMax founder &amp; CEO Yan Junjie expressed, &#34;For the first time, I feel that the mountain is not insurmountable.&#34;
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2e8b04b3-5e48-4b03-9d71-09c3765c5428.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2e8b04b3-5e48-4b03-9d71-09c3765c5428.png"/>
<h2>MiniMax-M1: Hybrid Expert Architecture and Linear Attention Mechanism</h2>
MiniMax-M1 adopts a hybrid expert (MoE) architecture and linear attention mechanism (Lightning Attention), which is a direct challenge to the computational bottlenecks of traditional Transformer architectures.
&#34;This design theoretically allows for efficient extension of inference length to hundreds of thousands of tokens,&#34; MiniMax stated, adding that it also brings a significant reduction in computational costs, &#34;this feature gives us a substantial computational efficiency advantage during both training and inference.&#34;
The model has a total of 456 billion parameters, with 45.9 billion parameters activated per token, supporting context inputs of up to 1 million tokens— this number is 8 times that of DeepSeek R1, tied for the highest in the industry with Google Gemini 2.5 Pro.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/7023e4b0-ec9a-4344-8e8f-d6ad05938245.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/7023e4b0-ec9a-4344-8e8f-d6ad05938245.png"/>
In tests across 17 mainstream evaluation sets, M1 achieved over 55% on the software engineering capability test SWE-bench. Although it did not reach the level of top overseas models, it surpassed DeepSeek-R1 and similar products from Alibaba and ByteDance. In long context understanding tasks, M1 comprehensively outperformed all open-source models in three benchmark tests, only slightly trailing behind Gemini 2.5 Pro, ranking second globally
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/37db4f0d-d61d-4145-8022-20a4914a7c8a.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/37db4f0d-d61d-4145-8022-20a4914a7c8a.png"/>
<h2>Cost Revolution: A Reinforcement Learning Experiment Costing 3.8 Million RMB</h2>
MiniMax claims that the entire reinforcement learning process of M1 only used 512 NVIDIA H800 GPUs for three weeks, with a rental cost of $537,400 (approximately 3.8 million RMB). This cost control &#34;is an order of magnitude less than initially expected.&#34;
The company has also developed a new reinforcement learning algorithm called CISPO, which achieved a twofold acceleration over the DAPO algorithm recently proposed by ByteDance in the mathematical testing benchmark AIME, requiring only 50% of the training steps to achieve the same performance.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/39925c09-0e70-4b51-9e4a-11d95a6c1bd5.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/39925c09-0e70-4b51-9e4a-11d95a6c1bd5.png"/>
Compared to DeepSeek R1, M1 consumed less than 50% of the computing power when generating 64K tokens; at 100K tokens, it was only 25%.
<h2>Tiered Pricing Strategy! MiniMax Has More Updates</h2>
Currently, MiniMax-M1 has been open-sourced and is available for free upgrades on the MiniMax APP and Web version. In terms of API pricing, MiniMax adopts the same &#34;tiered pricing&#34; strategy as ByteDance's Doubao 1.6.
In the input length ranges of 0-32k and 32k-128k, M1's pricing is more cost-effective compared to DeepSeek-R1 (4 RMB per million tokens for input, 16 RMB per million tokens for output). For the longest input range of 128k-1M, the DeepSeek model does not even support this length.
This pricing strategy makes M1 another &#34;price killer&#34; following Doubao, with developers rating it as the &#34;new king of cost performance.&#34;
<h2>The Survival Game of the &#34;AI Six Dragons&#34;</h2>
As one of the &#34;AI Six Dragons&#34; supported by Tencent and Alibaba, MiniMax continues to insist on basic research. MiniMax founder and CEO Yan Junjie stated, &#34;For the first time, I feel that the mountain is not insurmountable.&#34;
According to Sohu Technology, M1 is just the first product released during the company's five-day launch week, with subsequent releases of intelligent agent applications and more updates in video, music, and other model areas.
MiniMax believes that M1's efficient architecture will have unique advantages in future intelligent agent applications. &#34;Future intelligent agents will need to reason over dozens to hundreds of rounds while integrating long contextual information from different sources,&#34; the company stated. Currently, MiniMax is conducting internal testing of intelligent agent applications overseas, focusing on capabilities such as coding and multimodal functions

NVDA

AI startup MiniMax has released its first inference model M1, which was trained for three weeks using 512 NVIDIA H800 GPUs, with a rental cost of $537,400. In multiple benchmark tests, M1 surpassed DeepSeek's latest R1-0528 model, requiring only 25% of the computational resources used by DeepSeek to generate 100K tokens

- MiniMax has launched its first inference model, M1, which reportedly outperforms domestic closed-source models and rivals leading international models, with a training cost of approximately $53,740.  
- The M1 model utilizes a hybrid expert architecture and linear attention mechanism, achieving significant efficiency, consuming less than 50% of the computational power of DeepSeek R1 for generating 64K tokens.  
- MiniMax's innovative pricing strategy positions M1 as a cost-effective alternative in the market, enhancing its competitive edge among AI models supported by major companies like Tencent and Alibaba.

Wallstreetcn

NVIDIA

GraniteShares 2x Long NVDA Daily ETF

OpenAI

Alphabet - C

Venture

TENCENT

Alphabet

YieldMax NVDA Option Income Strategy ETF

Sohu.com

The 20 billion AI unicorn strikes back, MiniMax's first inference model rivals DeepSeeK, with computing costs of only 530,000 USD